This is the official code for "Wide Two-Layer Networks can Learn from Adversarial Perturbations" [S. Kumano et al., NeurIPS24].

All generated data, including synthetic datasets, model parameters, and adversarial examples, can be downloaded from here or here (48GB).

Setup

docker-compose -f "docker/docker-compose.yaml" up -d --build 

Run

bash/train.sh <gpu_id: int>
bash/cossim.sh <gpu_id: int>
bash/map.sh <gpu_id: int>